Corrosion is a major concern for aging steel bridges because progressive material loss can affect their long-term serviceability and, in severe cases, structural safety. Visual inspection remains indispensable for assessing bridge condition, but inspectors must evaluate corrosion under widely varying conditions, including differences in lighting, coating color, shadows, surface deposits, structural geometry, and accessibility. These factors can make consistent corrosion assessment challenging.
Deep learning has increasingly been used to automate corrosion detection in infrastructure images. Most existing systems, however, treat the problem as a binary task: each pixel is classified simply as either "corrosion" or "background." Although this is useful for determining the location and area of corrosion, it does not distinguish between visually different forms of deterioration. Broad corrosion over an exposed steel surface, corrosion concentrated around a joint, and deterioration associated with coating failure may look different and may call for different follow-up observations during an inspection. Existing approaches also rely mainly on measures of how well predicted corrosion regions overlap with reference images, while the accuracy of irregular corrosion boundaries has received less attention.
A Saitama University research team comprising PhD student Sal Saad Al Deen Taher and Associate Professor Ji Dang from the Department of Civil and Environmental Engineering therefore set out to develop a more informative way of representing visible corrosion in steel bridge inspection images. Using field images obtained from Japanese steel bridge inspection reports, the researchers constructed a binary corrosion dataset containing 1,960 images and a categorized dataset containing 1,738 images. They evaluated both convolutional neural network- and transformer-based deep learning models and introduced a two-level framework in which Level 1 identifies the presence and spatial extent of visible corrosion, while Level 2 classifies corrosion pixels into four inspection-oriented visual categories: Uniform, Crevice, Underfilm, and Other Localized corrosion. In this way, the team successfully developed a framework that provides complementary information about where corrosion occurs, what visual pattern it exhibits, and how accurately its boundaries are represented.
The study was published online in Computer-Aided Civil and Infrastructure Engineering on August 5, 2026.
Key findings of the study include:
- For Level 1 corrosion-coverage detection, the best-performing configuration, SegFormer-B2 with boundary-aware supervision, achieved a mean Intersection over Union (mIoU) of 84.30%, a corrosion IoU of 73.39%, and a corrosion F1-score of 84.65%. Mask2Former-Small achieved the highest Boundary-F1 score of 69.10%, showing that the model with the best overall region overlap is not necessarily the one that most accurately reproduces corrosion boundaries.
- For Level 2 visual corrosion-pattern classification, the adopted SegFormer-B5 configuration achieved a five-class mIoU of 60.45% and a mean F1-score of 67.81% across the four corrosion categories.
- The four-category visual annotation system also showed strong reproducibility. In a blinded second-annotator audit, overall agreement reached 92.81%, with a Cohen's κ of 0.892 when uncertain responses were counted as disagreements.
- Importantly, the two levels are designed to complement rather than replace one another. Level 1 provides the primary estimate of corrosion presence and coverage, while Level 2 provides supplementary information that can help inspectors decide what types of field observations and measurements should receive closer attention.
"Conventional AI-based corrosion segmentation can tell us where corrosion is visible, but inspectors often need more information than a simple yes-or-no corrosion map," says Taher. "Our framework adds another layer of information by distinguishing visually different corrosion patterns at the pixel level. This could help connect automated image analysis more directly with the way inspectors decide where to look more closely and what should be checked in the field."
For example, a Uniform corrosion prediction can direct attention toward the overall extent of surface deterioration and remaining-thickness measurements, while a Crevice prediction can highlight joints, contact surfaces, or fastener regions that warrant closer examination. Underfilm predictions can indicate areas where damaged coatings may need to be probed or inspected more closely, while Other Localized corrosion can direct attention toward localized pits, deposits, deteriorated fasteners, or other concentrated deterioration. The system is intended to guide field verification and measurement—not to make autonomous decisions about repairs, maintenance priority, or structural capacity.
Looking five to ten years ahead, the approach could become particularly useful as bridge inspection increasingly incorporates digital imaging, mobile devices, unmanned aerial vehicles, and automated infrastructure-management systems. Future studies will need to test the framework on independently collected bridge datasets, under different viewing angles, lighting conditions, image resolutions, and geographical environments, as well as on mobile and edge-computing hardware. The researchers also envision combining corrosion-coverage maps and visual-category information with inspection histories, structural information, and physical measurements in bridge-management systems.
"In the longer term, we hope this type of technology can become a practical assistant for engineers rather than an autonomous replacement for them," Taher says. "If image-based corrosion information can be combined with inspection records and physical measurements, inspectors could review bridge conditions in a more systematic and traceable way. That could help infrastructure owners make better use of inspection data and focus engineering attention where it is most needed."
The researchers emphasize that the present framework evaluates visible corrosion appearances, rather than directly measuring section loss, residual load-carrying capacity, deterioration rate, or maintenance priority. Such engineering decisions will continue to require physical measurements, inspection records, and expert judgment.
